import gradio as gr
import numpy as np
from PIL import Image
import cv2
from insightface.app import FaceAnalysis
from huggingface_hub import snapshot_download
import time
import subprocess
import os
# --- Configuration ---
SECURITYLEVELS = ["128", "196", "256"]
FRMODELS = ["AuraFace-v1"]
EXAMPLE_IMAGES_ENROLL = ['./VGGFace2/n000001/0002_01.jpg', './VGGFace2/n000149/0002_01.jpg', './VGGFace2/n000082/0001_02.jpg', './VGGFace2/n000148/0014_01.jpg']
EXAMPLE_IMAGES_AUTH = ['./VGGFace2/n000001/0013_01.jpg', './VGGFace2/n000149/0019_01.jpg', './VGGFace2/n000082/0003_03.jpg', './VGGFace2/n000148/0043_01.jpg']
# --- Global Variables ---
face_app = None
DB_SUBJECT_COUNT = 1
ENROLLED_SEARCH_IMAGES = []
# --- Helper Functions ---
def initialize_face_app():
"""Initializes the FaceAnalysis model."""
global face_app
if face_app is None:
print("Initializing FaceAnalysis model...")
snapshot_download("fal/AuraFace-v1", local_dir="./models/auraface")
face_app = FaceAnalysis(name="auraface", providers=["CPUExecutionProvider"], root=".")
face_app.prepare(ctx_id=0, det_size=(128, 128))
print("FaceAnalysis model initialized.")
return face_app
def run_binary(bin_path, *args):
"""Runs a compiled binary file and returns the result."""
if not os.path.isfile(bin_path):
raise gr.Error(f"Error: Compiled binary not found at {bin_path}")
command = [bin_path] + [str(arg) for arg in args]
print(f"Running command: {' '.join(command)}")
try:
os.chmod(bin_path, 0o755)
start_time = time.time()
result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=True)
duration = time.time() - start_time
print(f"Binary execution successful. Duration: {duration:.2f}s")
return result.stdout, duration
except subprocess.CalledProcessError as e:
print(f"Error executing binary: {e.stderr}")
raise gr.Error(f"Execution failed: {e.stderr}")
except Exception as e:
print(f"An unexpected error occurred: {e}")
raise gr.Error(f"An unexpected error occurred: {str(e)}")
def extract_embedding(image_path, mode=None):
"""Extracts face embedding from an image path."""
if image_path is None:
raise gr.Error("Please upload or select an image first.")
app = initialize_face_app()
try:
pil_image = Image.open(image_path).convert("RGB")
except Exception as e:
raise gr.Error(f"Failed to open or read image file: {e}")
cv2_image = np.array(pil_image)
cv2_image = cv2_image[:, :, ::-1]
faces = app.get(cv2_image)
if not faces:
raise gr.Error("No face detected. Please try another image.")
embedding = faces[0].normed_embedding
if mode:
# For 1:1 recognition, save to the respective binary folder
if mode in ["enroll", "auth"]:
emb_path = f'./{mode}-emb.txt'
# For 1:N search, create a subject-specific path in the search folder
else: # search_enroll, search_auth
if "VGGFace2" in image_path:
subject = image_path.split('/')[-2]
else:
subject = 'uploadedSubj'
os.makedirs(f'./embeddings/{subject}', exist_ok=True)
emb_path = f'./embeddings/{subject}/{mode}-emb.txt'
np.savetxt(emb_path, embedding.reshape(1, -1), fmt="%.6f", delimiter=',')
return embedding.tolist(), emb_path
return embedding.tolist()
# --- UI Components ---
def create_image_selection_ui(label, gallery_images):
with gr.Group():
gr.HTML(f'
')
image_state = gr.State()
image_display = gr.Image(type="filepath", label="Selected Image", interactive=False)
with gr.Tabs():
with gr.TabItem("Upload"):
image_upload = gr.Image(type="filepath", label=f"Upload Image")
with gr.TabItem("Select from Gallery"):
image_gallery = gr.Gallery(value=gallery_images, columns=4, height="auto", object_fit="contain")
# Event handlers that directly update both the hidden state and the visible display
def on_select(evt: gr.SelectData):
selected_image = gallery_images[evt.index] # Get the actual image path from the gallery list
return selected_image, selected_image
def on_upload(filepath):
return filepath, filepath
image_upload.change(on_upload, inputs=image_upload, outputs=[image_state, image_display])
image_gallery.select(on_select, None, outputs=[image_state, image_display])
return image_state
# --- UI Styling and Theming ---
css = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');
:root {
--background: #EEEEEC; --background-alt: #EEEEEC; --card-bg: #FFFFFF; --card-bg-alt: rgba(255, 208, 134, 0.3);
--foreground: #222; --foreground-muted: #333; --accent-orange: rgb(255, 208, 134);
--accent-gradient: linear-gradient(90deg, var(--accent-orange) 0%, #333333 100%);
--font-sans: 'Inter', Arial, Helvetica, sans-serif; --gray-333: #333333;
}
body, .gradio-container { background: var(--background); color: var(--foreground); font-family: var(--font-sans); font-size: 16px; line-height: 1.6; }
.main-header { padding: 1rem; text-align: center; margin-bottom: 2rem; background: var(--gray-333); color: var(--background); border-radius: 15px; }
.main-header h1 { font-size: 2.5rem; font-weight: 700; color: var(--accent-orange); margin:0; }
.main-header p { font-size: 1.1rem; opacity: 0.9; margin: 0.5rem 0 0 0; }
.main-header a { color: var(--background); text-decoration: none; background: transparent; padding: 0.6rem 1.5rem; border-radius: 25px; border: 1px solid var(--accent-orange); font-weight: 500; transition: all 0.3s ease; display: inline-block; margin-top: 1rem; }
.main-header a:hover { background: var(--accent-orange); color: var(--gray-333); }
.section-header { text-align: center; margin: 2rem 0; padding: 0 1rem; }
.section-header h1 { color: var(--foreground); font-size: 2.2rem; font-weight: 600; margin-bottom: 0.5rem; }
.section-header h2 { color: var(--foreground-muted); font-size: 1.5rem; font-weight: 400; margin: 0; }
.narrative-section { background: var(--card-bg); border-top: 4px solid var(--accent-orange); padding: 2rem; margin: 1.5rem 0; border-radius: 12px; box-shadow: 0 4px 15px rgba(0,0,0,0.05); }
.narrative-header { color: var(--foreground); margin: 0 0 1.5rem 0; font-size: 1.8rem; font-weight: 600; }
.step-header { color: var(--foreground); margin: 0 0 1.5rem 0; font-size: 1.3rem; font-weight: 600; }
.info-card { background: var(--card-bg-alt); border: 1px solid var(--accent-orange); border-radius: 12px; padding: 1.5rem; margin: 1.5rem 0; }
.info-card h3 { color: var(--foreground); margin: 0 0 1rem 0; font-size: 1.3rem; font-weight: 600; }
.info-card p { margin: 0 0 1rem 0; color: var(--foreground-muted); line-height: 1.6; }
.warning-card { background: #ffebee; border: 1px solid #c62828; border-radius: 12px; padding: 1.5rem; margin: 1.5rem 0; }
.warning-card h3 { color: #c62828; margin: 0 0 1rem 0; font-size: 1.3rem; font-weight: 600; }
.warning-card p { margin: 0; color: #424242; line-height: 1.6; }
.result-container { padding: 2rem; border-radius: 15px; text-align: center; margin-top: 1rem; color: white; }
.result-container h2 { margin: 0 0 0.5rem 0; font-size: 2rem; font-weight: 600; color: white; }
.result-container p { margin: 0; opacity: 0.95; font-size: 1rem; }
.match-verified { background: linear-gradient(135deg, #4caf50 0%, #45a049 100%); }
.no-match { background: linear-gradient(135deg, #f44336 0%, #d32f2f 100%); }
.icon-lock { font-size: 4rem; margin: 1rem; }
.status-text { font-size: 1.1rem; color: var(--foreground-muted); margin-top: 1rem; }
"""
# --- Gradio UI Definition ---
with gr.Blocks(css=css) as demo:
# --- Header ---
gr.HTML("""
Suraksh AI
The Future of Secure Biometrics
π Visit Our Website
""")
# --- Key Generation on Load ---
# Generate keys once for each demo when the app starts up
demo.load(lambda: run_binary("./bin/genKeys.bin", "128", "genkeys"), None, None)
demo.load(lambda: run_binary("./bin/search.bin", "128", "genkeys"), None, None)
# --- Main Tabs for Demo Mode ---
with gr.Tabs() as mode_tabs:
# --- 1:1 Recognition Demo ---
with gr.TabItem("ποΈ Face Recognition (1:1)"):
gr.HTML("""""")
with gr.Tabs():
# --- Vulnerable System Tab ---
with gr.TabItem("π¨ The Vulnerable System"):
with gr.Group(elem_classes="narrative-section"):
gr.HTML('')
gr.HTML("""β οΈ Your Biometric Data is Exposed!
Most systems handle biometric data in plaintext. This means your facial embeddingβa digital map of your faceβcan be stolen and used to reconstruct your image, creating a major privacy risk.
""")
with gr.Column():
gr.HTML('')
gr.Image(value=EXAMPLE_IMAGES_ENROLL[2], label="Original Face", interactive=False, show_label=False, container=False)
gr.HTML('An attacker breaches the system and steals the stored facial embedding. Click the button to simulate this theft.
')
extract_btn = gr.Button("π± Steal Biometric Data", variant="primary")
with gr.Group(visible=False) as stolen_data_group:
feature_output = gr.JSON(label="Stolen Feature Vector (Face Embedding)")
gr.HTML('Now, the attacker uses the stolen features to create a reconstruction of the face, completely compromising the user\'s privacy.
')
reconstruct_btn = gr.Button("π Reconstruct Face from Stolen Data", variant="stop")
with gr.Group(visible=False) as reconstructed_image_group:
reconstructed_output = gr.Image(label="Reconstructed Face", interactive=False, show_label=False)
def extract_and_reveal(image_path):
embedding = extract_embedding(image_path)
feature_json = {"embedding": embedding}
return {
feature_output: feature_json,
stolen_data_group: gr.update(visible=True),
extract_btn: gr.update(value="Data Stolen!", interactive=False)
}
def show_reconstruction():
reconstructed_image_path = "./static/reconstructed.png"
return {
reconstructed_output: reconstructed_image_path,
reconstructed_image_group: gr.update(visible=True),
reconstruct_btn: gr.update(interactive=False)
}
extract_btn.click(
fn=extract_and_reveal,
inputs=gr.State(EXAMPLE_IMAGES_ENROLL[0]),
outputs=[feature_output, stolen_data_group, extract_btn]
)
reconstruct_btn.click(
fn=show_reconstruction,
inputs=None,
outputs=[reconstructed_output, reconstructed_image_group, reconstruct_btn]
)
# --- Secure System Tab ---
with gr.TabItem("β
The Suraksh.AI Solution"):
with gr.Group(elem_classes="narrative-section"):
gr.HTML('')
gr.HTML("""The Locked Box Analogy
With Suraksh.AI, your biometric data is encrypted inside a "locked box" before it ever leaves your device. We can perform the verification on the encrypted data without ever seeing your real face. It's mathematically impossible for us to decrypt it.
""")
with gr.Row():
with gr.Column():
rec_ref_img = create_image_selection_ui("1. Provide Reference Image", EXAMPLE_IMAGES_ENROLL)
with gr.Group(visible=False) as rec_ref_features_group:
rec_ref_raw_features = gr.JSON(label="Raw Features (Plaintext)")
rec_ref_encrypted_features = gr.Textbox(label="Encrypted Features (Ciphertext)", interactive=False, lines=5)
with gr.Column():
rec_probe_img = create_image_selection_ui("2. Provide Probe Image", EXAMPLE_IMAGES_AUTH)
with gr.Group(visible=False) as rec_probe_features_group:
rec_probe_raw_features = gr.JSON(label="Raw Features (Plaintext)")
rec_probe_encrypted_features = gr.Textbox(label="Encrypted Features (Ciphertext)", interactive=False, lines=5)
with gr.Accordion("Advanced Settings", open=False):
rec_threshold = gr.Slider(-512*5, 512*5, value=133, label="Match Strictness", info="A higher value means a stricter match is required.")
rec_sec_level = gr.Dropdown(SECURITYLEVELS, value="128", label="Security Level")
rec_run_btn = gr.Button("π Perform Secure 1:1 Match", variant="primary", size="lg")
rec_status = gr.HTML(elem_classes="status-text")
rec_result = gr.HTML()
def secure_recognition_flow(ref_img, probe_img, threshold, sec_level):
# Reset UI
yield "Initializing...", "", gr.update(visible=False), None, None, gr.update(visible=False), None, None
# Process Reference Image
yield "Extracting reference features...", "", gr.update(visible=False), None, None, gr.update(visible=False), None, None
ref_emb, _ = extract_embedding(ref_img, "enroll")
yield "Encrypting reference features...", "", gr.update(visible=True), {"embedding": ref_emb}, None, gr.update(visible=False), None, None
run_binary("./bin/encReference.bin", sec_level, "encrypt")
ref_ciphertext, _ = run_binary("./bin/encReference.bin", sec_level, "print")
# Process Probe Image
yield "β
Reference Encrypted. Extracting probe features...", "", gr.update(visible=True), {"embedding": ref_emb}, ref_ciphertext, gr.update(visible=False), None, None
probe_emb, _ = extract_embedding(probe_img, "auth")
yield "Encrypting probe features...", "", gr.update(visible=True), {"embedding": ref_emb}, ref_ciphertext, gr.update(visible=True), {"embedding": probe_emb}, None
run_binary("./bin/encProbe.bin", sec_level, "encrypt")
probe_ciphertext, _ = run_binary("./bin/encProbe.bin", sec_level, "print")
# Perform Match
yield "β
Probe Encrypted. Performing Secure Match...", "", gr.update(visible=True), {"embedding": ref_emb}, ref_ciphertext, gr.update(visible=True), {"embedding": probe_emb}, probe_ciphertext
run_binary("./bin/recDecision.bin", sec_level, "decision", threshold)
yield "β
Match Computed. Decrypting Result...", "", gr.update(visible=True), {"embedding": ref_emb}, ref_ciphertext, gr.update(visible=True), {"embedding": probe_emb}, probe_ciphertext
output, _ = run_binary("./bin/decDecision.bin", sec_level, "decision")
if output.strip().lower() == "match":
result_html = f"""β
MATCH VERIFIED
Identity successfully confirmed under FHE.
"""
else:
result_html = f"""β NO MATCH
Identity verification failed.
"""
yield "Done!", result_html, gr.update(visible=True), {"embedding": ref_emb}, ref_ciphertext, gr.update(visible=True), {"embedding": probe_emb}, probe_ciphertext
rec_run_btn.click(
fn=secure_recognition_flow,
inputs=[rec_ref_img, rec_probe_img, rec_threshold, rec_sec_level],
outputs=[rec_status, rec_result, rec_ref_features_group, rec_ref_raw_features, rec_ref_encrypted_features, rec_probe_features_group, rec_probe_raw_features, rec_probe_encrypted_features]
)
# --- 1:N Search Demo ---
with gr.TabItem("π Face Search (1:N)"):
gr.HTML("""""")
with gr.Tabs():
# --- Secure System Tab ---
with gr.TabItem("β
The Suraksh.AI Solution"):
with gr.Group(elem_classes="narrative-section"):
gr.HTML('')
gr.HTML("""From Verification to Identification
This demo shows how FHE can be used to search for a person in a database without ever decrypting the database itself. This is ideal for large-scale, privacy-preserving identification systems.
""")
with gr.Row():
with gr.Column():
gr.HTML('')
search_enroll_img = create_image_selection_ui("Select Image to Enroll", EXAMPLE_IMAGES_ENROLL)
search_enroll_btn = gr.Button("β Encrypt & Add to Database", variant="secondary")
with gr.Group(visible=False) as enroll_features_group:
enroll_raw_features = gr.JSON(label="Raw Features (Plaintext)")
enroll_encrypted_features = gr.Textbox(label="Encrypted Features (Ciphertext)", interactive=False, lines=5)
search_enroll_status = gr.HTML()
with gr.Column():
gr.HTML('')
search_probe_img = create_image_selection_ui("Select Image to Search", EXAMPLE_IMAGES_AUTH)
search_run_btn = gr.Button("π Perform Secure 1:N Search", variant="primary", size="lg")
with gr.Group(visible=False) as search_features_group:
search_raw_features = gr.JSON(label="Raw Features (Plaintext)")
search_encrypted_features = gr.Textbox(label="Encrypted Features (Ciphertext)", interactive=False, lines=5)
search_status = gr.HTML(elem_classes="status-text")
search_result = gr.HTML()
search_result_image = gr.Image(label="Found Reference Image", interactive=False, visible=False)
with gr.Accordion("Advanced Settings", open=False):
search_threshold = gr.Slider(-512*5, 512*5, value=133, label="Match Strictness")
search_sec_level = gr.Dropdown(SECURITYLEVELS, value="128", label="Security Level")
def secure_enroll_flow(image, sec_level):
global DB_SUBJECT_COUNT, ENROLLED_SEARCH_IMAGES
if image is None: raise gr.Error("Please provide an image to enroll.")
current_id = DB_SUBJECT_COUNT
yield "Extracting features...", gr.update(visible=False), None, None
embedding, emb_path = extract_embedding(image, "search_enroll")
yield "Encrypting features...", gr.update(visible=True), {"embedding": embedding}, None
run_binary("./bin/search.bin", sec_level, "encRef", emb_path, current_id)
ciphertext, _ = run_binary("./bin/search.bin", sec_level, "printVectorCipher", "encRef", "print")
yield "Adding to secure database...", gr.update(visible=True), {"embedding": embedding}, ciphertext
run_binary("./bin/search.bin", sec_level, "addRef")
ENROLLED_SEARCH_IMAGES.append(image)
DB_SUBJECT_COUNT += 1
yield f"β
Subject with ID {current_id} added. Total subjects: {DB_SUBJECT_COUNT - 1}.", gr.update(visible=True), {"embedding": embedding}, ciphertext
def secure_search_flow(image, threshold, sec_level):
global ENROLLED_SEARCH_IMAGES
if image is None: raise gr.Error("Please provide an image to search.")
yield "Extracting probe features...", "", gr.update(visible=False), None, None, gr.update(visible=False)
embedding, emb_path = extract_embedding(image, "search_auth")
yield "Encrypting probe features...", "", gr.update(visible=True), {"embedding": embedding}, None, gr.update(visible=False)
run_binary("./bin/search.bin", sec_level, "encProbe", emb_path)
ciphertext, _ = run_binary("./bin/search.bin", sec_level, "printProbe", "print")
yield "β
Probe encrypted. Searching database...", "", gr.update(visible=True), {"embedding": embedding}, ciphertext, gr.update(visible=False)
run_binary("./bin/search.bin", sec_level, "search")
yield "β
Search complete. Decrypting results...", "", gr.update(visible=True), {"embedding": embedding}, ciphertext, gr.update(visible=False)
# output, _ = run_binary("./bin/search.bin", sec_level, "decDecisionClear", threshold)
output, _ = run_binary("./bin/search.bin", sec_level, "decScoreClear", threshold)
print(f"Search binary output: >>>{output}<<<")
lines = output.strip().split('\n')
decision = lines[0].lower()
if decision == "found":
try:
# Assuming output is "found\nID: "
found_id_line = next(line for line in lines if "id:" in line.lower())
found_id = int(found_id_line.split(':')[1].strip())
# Binary ID is 1-based, list is 0-based
found_image_path = ENROLLED_SEARCH_IMAGES[found_id - 1]
result_html = f"""β
SUBJECT FOUND
The subject was successfully found in the database with ID {found_id}.
"""
result_image_update = gr.update(value=found_image_path, visible=True)
except (StopIteration, IndexError, ValueError):
result_html = f"""β
SUBJECT FOUND
Could not parse ID from binary output: {output.strip()}
"""
result_image_update = gr.update(visible=False)
else:
result_html = """β NOT FOUND
The subject was not found in the database.
"""
result_image_update = gr.update(visible=False)
yield "Done!", result_html, gr.update(visible=True), {"embedding": embedding}, ciphertext, result_image_update
search_enroll_btn.click(
fn=secure_enroll_flow,
inputs=[search_enroll_img, search_sec_level],
outputs=[search_enroll_status, enroll_features_group, enroll_raw_features, enroll_encrypted_features]
)
search_run_btn.click(
fn=secure_search_flow,
inputs=[search_probe_img, search_threshold, search_sec_level],
outputs=[search_status, search_result, search_features_group, search_raw_features, search_encrypted_features, search_result_image]
)
# --- Launch the Application ---
if __name__ == "__main__":
demo.launch()